Quickstart

Inspect a supported checkpoint, deploy it, and predict an action chunk.

Choose a model

1. Get actions

   Run this on the robot computer with the deployment ID from the deploy command. The client prints the deployed policy’s resolved input keys, shapes, and dtypes before predicting. The sample sends dummy arrays for illustration; replace them with your camera and state arrays, keeping each inspected key, shape, and dtype unchanged.

   first_prediction.py

   import numpy as np
   import servo
   
   sv = servo.Servo()
   deployment = sv.deployments.get(input("Deployment ID: "))
   policy = deployment.policy()
   print("Inputs required by this deployment:")
   for feature in policy.expected_inputs:
       print(f"  {feature.name}: shape={feature.shape}, dtype={feature.dtype}")
   np.random.seed(0)
   inputs = {
       feature.name: (
           np.random.randint(0, 256, feature.shape, dtype=np.uint8)
           if feature.dtype == "uint8"
           else np.array(np.random.standard_normal(feature.shape), dtype=feature.dtype)
       )
       for feature in policy.expected_inputs
   }
   with sv.session(policy, instruction="pick up the object") as session:
       actions = session.predict(inputs=inputs).actions
   print(f"Actions: {len(actions)} steps x {len(actions[0])} values")
   

   Download and run the sample on the robot computer.

   robot computer

   curl -fsSL https://api.orchestrallabs.ai/docs/downloads/first_prediction.py -o first_prediction.py
   python first_prediction.py
   
2. Stop the deployment

   Stop the deployment when finished.

   your computer

   servo deployment stop DEPLOYMENT_ID --wait